PaperCraft

From rough idea to accepted paper at top CS conferences.

Not just another paper-writing tool — a full-pipeline AI research partner.

🌟 If PaperCraft helps your research, please consider giving it a star — it helps others discover it!

npx paper-craft-ai

English · 中文说明 · The 12 Skills · The 17 Rules · Changelog · Cross-Tool

🎯 The Problem

Every AI coding assistant does this:

| ❌ What Happens Without PaperCraft | ✅ What PaperCraft Does | |:-:|:-:| | "Help me search papers" → returns 100 random results | /paper-search — 4-layer progressive search with quality filtering and citation traversal | | Writes without checking if the idea is novel | /paper-idea — novelty check against existing work before you invest time | | Designs experiments needing 8×A100 when you have 1×3090 | /paper-plan — asks about your GPU first, designs within constraints | | "Write me a method section" → generic template output | /paper-write — conference-specific style, claim-evidence alignment | | Training crashes with NaN loss, no idea why | /paper-train — auto-derive params from your GPU, debug with decision trees | | Submits paper, gets reviewer 2'd | /paper-review — simulate 3 reviewers including a skeptical one, with conference rubric | | Doesn't know where to start | /paper-launch — one command from raw ideas to complete project plan |

PaperCraft encodes the entire CS research pipeline — from literature search to post-submission rebuttal — into 12 skills and 17 rules that make AI a real research partner, not just a writing assistant.

⚡ Quick Start

# Install with npx (recommended, cross-platform)
npx paper-craft-ai

# Or with curl (macOS/Linux)
curl -fsSL https://raw.githubusercontent.com/charlotte-12s/paper-craft/main/install.sh | bash

# Or clone and run
git clone https://github.com/charlotte-12s/paper-craft.git
cd paper-craft
node cli.js

# Install only what you need
npx paper-craft-ai --bundle search-only   # Literature search only
npx paper-craft-ai --bundle write-only    # Writing pipeline (story + write + review)
npx paper-craft-ai --bundle launch-only   # One-command launcher only

# Install for specific AI tools
npx paper-craft-ai --tool claude-code --tool cursor

🔄 The Full Pipeline

 💡 Launch       🔍 Search       💭 Idea         📋 Plan         📊 Data         🔧 Env
paper-launch → paper-search → paper-idea → paper-plan → paper-data → paper-env
                                                                    ↓
 🏋️ Train       📖 Story        ✍️ Write        🔎 Review       💬 Rebuttal     ✅ Audit
paper-train → paper-story → paper-write → paper-review → paper-rebuttal → paper-audit

Each skill is independently callable — enter at any point and loop back as needed. Real research is iterative; PaperCraft supports that.

🛠️ The 12 Skills

| Skill | Command | Stage | What It Does | |-------|---------|:-----:|-------------| | Research Launcher | /paper-launch | 🚀 Full | One-command: ideas → literature search → novelty check → experiment plan → project package | | Paper Search | /paper-search | 🔍 Find | 4-layer progressive search (core/supporting/validation/researcher) with snowball citation traversal | | Idea Check | /paper-idea | 💭 Validate | Find innovation gaps, novelty check decision tree, risk assessment | | Experiment Plan | /paper-plan | 📋 Design | Resource-aware design (asks GPU first), content-driven ablation planning | | Data Pipeline | /paper-data | 📊 Prepare | Dataset selection, contamination check, one-click processing pipeline | | Environment Setup | /paper-env | 🔧 Setup | One-shot setup with version compatibility matrix (CUDA/PyTorch/Python) | | Training Config | /paper-train | 🏋️ Train | Auto-derive params from GPU, NaN/OOM debug decision trees, result analysis | | Narrative Story | /paper-story | 📖 Craft | 6 CS research story patterns with narrative arc construction | | Paper Writing | /paper-write | ✍️ Write | Section-by-section with conference-specific style guides and revision | | Review Simulation | /paper-review | 🔎 Review | Simulate 2-3 reviewers (including skeptical one) with conference rubric | | Rebuttal Draft | /paper-rebuttal | 💬 Respond | Parse reviews, draft evidence-based responses, plan resubmission | | Pipeline Audit | /paper-audit | ✅ Audit | 50+ item checklist across literature/novelty/experiments/writing/reproducibility |

Skill Detail

The "wheelchair mode" — you provide ideas + compute, AI does the rest:

1. Conference Recommendation — Analyze your direction, recommend 2-3 target venues from 16 CCF-A profiles
2. Idea Collection — Collect 2-5 research ideas + compute resources (GPU type/count/memory)
3. Idea Evaluation — Auto-search literature, score on Novelty/Feasibility/Conference Match/Acceptance Probability
4. Auto-Search — After you pick an idea, auto-run paper-search + paper-idea + find open-source code/datasets
5. Project Package — Generate directory skeleton, config files, run scripts, LaTeX template, experiment checklist

Human checkpoints at: conference selection → idea selection → plan confirmation.

10-step progressive search with quality and relevance filtering:

1. Decompose Search Intent — Break direction into Core / Supporting / Validation layers
2. Construct Queries — Concept decomposition → synonym expansion → domain terminology → time range strategy
3. Core Layer Search — Google Scholar + arXiv + Semantic Scholar + DBLP, with quality (⭐⭐⭐/⭐⭐/⭐/❌) and relevance (🔴/🟡/🟢/⚪) labels
4. Snowball Expansion — Forward + backward citation traversal from 3-5 seed papers, 2-3 layers
5. Supporting Layer — Component-level work, identify differentiation opportunities
6. Validation Layer — Baselines and benchmarks from Papers with Code
7. Iterative Refinement — Too few → broaden; too many → narrow
8. Completeness Check — 3 consecutive rounds yield no new ⭐⭐⭐ papers → sufficient
9. Cross-Conference Analysis — If target conference selected, load profile and give conference-specific advice
10. Literature Report — Survey summary, innovation gaps, citation graph, Top 10 recommended reading

Includes: search-sources.md (4-layer source catalog) · query-construction.md (6-step methodology) · quality-filter.md (4-tier quality criteria) · relevance-scoring.md (decision tree) · citation-traversal.md (forward/backward strategy)

From GPU specs to publication-ready tables:

1. Auto-Derive Parameters — batch_size (max GPU memory), learning_rate (scale with batch), LoRA rank/alpha (by model size), warmup, weight_decay
2. Generate Configs — LLaMA-Factory YAML, DeepSpeed JSON, custom argparse scripts
3. Monitoring Guide — Normal/Warning/Critical thresholds for training loss, validation loss, LR, GPU utilization, gradient norm
4. Debug Mode — OOM → reduce batch/checkpointing/ZeRO; NaN loss → reduce LR/check data/reduce rank; Low accuracy → check data loading/model init
5. Results Analysis — Evaluation metrics, comparison tables (bold best, underline second), statistical significance tests, learning curves
6. Training Package — Configs + startup commands + monitoring guide + evaluation + LaTeX tables + matplotlib/TikZ figures

Includes: training-recipes.md (A100/A800/昇腾910B GPU-specific configs with VRAM estimates) · results-analysis.md (evaluation templates)

Choose a narrative arc that matches your contribution type:

| Pattern | Best When | Famous Examples | |---------|-----------|----------------| | Old Problem, New Paradigm | You solve an existing problem with a fundamentally different approach | Attention Is All You Need, Dropout as Bayesian | | Observation-Driven | You discover something surprising and build around it | ViT vs CNNs, LoRA | | Unified Framework | You unify scattered methods under one framework | SHAP, VAE | | Technical Breakthrough | You invent a new technique that enables new capabilities | Adam, FlashAttention | | Application-Driven | You bring a method to a high-impact domain | AlphaFold, BERT for Clinical | | Benchmark/Resource | You create a benchmark or dataset that enables future work | ImageNet, SuperGLUE |

Each pattern comes with a narrative template and CS paper examples with citations.

Includes: story-patterns.md (selection decision tree + detailed examples)

50+ item checklist across 6 dimensions:

| Dimension | Key Checks | |-----------|-----------| | Literature | All related work cited? No "first" claims without verification? Self-citation rate <15%? | | Novelty | Innovation points verified against literature? Differentiation from closest related work clear? | | Experiments | Baselines given same hyperparameter budget? Ablation covers every design choice? All benchmarks reported? | | Writing | Claim-evidence alignment table complete? No logic gaps in argument chain? Abstract matches content? | | Reproducibility | Random seeds reported? All hyperparameters specified? One-command reproduction script provided? | | Submission | Anonymization complete? LaTeX compiles without warnings? References formatted per conference? |

Includes: audit-checklist.md (complete 50+ item checklist with thresholds)

| Skill | Key Features | |-------|-------------| | /paper-launch | Conference recommendation → idea evaluation → auto-search → project package | | /paper-search | 4-layer search → quality/relevance filtering → snowball citations → literature report | | /paper-idea | Innovation gap mining → novelty decision tree → risk assessment | | /paper-plan | GPU-first design → open-source code/dataset search → content-driven ablation planning | | /paper-data | Dataset selection → contamination check → processing pipeline | | /paper-env | One-shot setup → CUDA/PyTorch compatibility matrix → verified stable combinations | | /paper-train | Auto-derive params → NaN/OOM debug → result tables and figures | | /paper-story | 6 story patterns → narrative arc → logic chain verification | | /paper-write | 8 section guides → conference-specific style → claim-evidence alignment | | /paper-review | 2-3 simulated reviewers → conference rubric → revision priority | | /paper-rebuttal | Review parsing → response patterns → venue matching strategy | | /paper-audit | 50+ item checklist → 6 dimensions → submission readiness |

🏛️ 16 CCF-A Conference Profiles

Each profile includes 7 sections: review weights, writing style, recent trends, reviewer concerns, anti-patterns, and success patterns — so your paper speaks the conference's language.

| Area | Conferences | |------|------------| | ML | ICML · ICLR · NeurIPS | | AI | AAAI · IJCAI | | CV | CVPR · ICCV · ECCV | | NLP | ACL · EMNLP | | DM | KDD · SIGIR · WWW · WSDM | | SE | ICSE | | Security | S&P |

Take NeurIPS as an example:

• Review Weights — Quality 30%, Clarity 20%, Originality 25%, Significance 25%
• Writing Style — Theoretical grounding expected; motivate with intuitive examples before formal definitions
• Recent Trends — Scaling laws, foundation model adaptations, mechanistic interpretability
• Reviewer Common Concerns — "How does this scale?", "Where is the theoretical justification?", "Why not just scale the baseline?"
• Anti-patterns — Pure empirical without theory, claims beyond what experiments support, ignoring computational cost
• Success Patterns — Clean theory + strong empirical, novel architecture with principled motivation, connecting seemingly unrelated fields

Every one of the 16 profiles is this specific.

📜 The 17 Research Rules

Rules that prevent the most common reasons papers get rejected:

| | Rule | Prevents | |:-:|------|:------------| | 1 | Novelty First — Confirm innovation before implementation | Building something that already exists | | 2 | Match Conference Taste — Adapt to target conference preferences | Using NeurIPS style for an AAAI submission | | 3 | Search Before Propose — Always search existing work first | Duplicating existing work | | 4 | Evidence Over Claims — Back every claim with data | Unsupported claims reviewers will attack | | 5 | Baseline Fairness — Compare against strong, recent baselines | Cherry-picking weak baselines | | 6 | Ablation Discipline — Verify every design choice | Adding modules without proof they help | | 7 | Open Source First — Find existing code before building | Reimplementing what already exists | | 8 | Reproducibility by Default — Record code + data + seeds | "Works on my machine" | | 9 | Writing Serves Story — Every section serves the narrative | Data dumps without narrative | | 10 | Review Anticipation — Preemptively address objections | Surprised by reviewer objections | | 11 | Consult Before Act — Human checkpoints at key decisions | AI running the entire pipeline alone | | 12 | Resource-Aware Design — Design experiments within compute limits | Experiments requiring 8×A100 when you have 1×3090 | | 13 | Prefer Reproducible Quality — Cite high-quality, reproducible papers | Citing non-reproducible papers to pad references | | 14 | Relevance Over Quantity — Filter search results tightly | 100 search results where 3 matter | | 15 | Explain Before Proceed — Always explain "why" and "what next" | Data without context or next steps | | 16 | Audit Before Submit — Mandatory pre-submission audit | Submitting with known errors | | 17 | Content-Driven Presentation — Figures/ablations from content, not templates | Generic figure templates that prove nothing |

💡 What Makes It Different?

Conference-Specific Adaptation

The same idea written differently for different venues. PaperCraft loads the target conference's profile and adapts:

• Writing style — NeurIPS values theoretical grounding; CVPR prioritizes visual results
• Experiment design — ICML expects ablation on every component; AAAI prefers breadth over depth
• Review rubric — Different weights on novelty/significance/technical quality per venue

Explain-Before-Proceed Pattern

Every output follows this format — no more data dumps without context:

📊 Result: What was done, what was found
💡 Explanation: Why this result, what it means for you
🎯 Action: What you need to do next

CS Research Standards Built In

• LaTeX double-column format (not Word)
• Algorithm pseudocode with \begin{algorithm}
• booktabs tables (bold best, underline second-best, no vertical lines)
• Color-blind friendly palettes (viridis/Set2, not red-green)
• Vector figures (PDF format, zoom-friendly)
• BibTeX + numeric citations

China-Specific Support

• A800 GPU and 昇腾910B training recipes
• CANN toolkit compatibility matrix
• MindSpore/PyTorch on Ascend setup guide
• Tsinghua/Aliyun PyPI mirrors
• HuggingFace mirrors (hf-mirror.com, modelscope.cn)

🔌 Cross-Tool Compatibility

Works with your favorite AI coding tool — auto-detected or specified manually:

| Tool | --tool flag | Auto-detection | |------|---------------|----------------| | Claude Code | claude-code | .claude/ directory | | Cursor | cursor | .cursor/ directory | | Codex CLI | codex | AGENTS.md file | | Gemini CLI | gemini | GEMINI.md file | | GitHub Copilot | copilot | .github/ directory | | Windsurf | windsurf | .windsurfrules file |

npx paper-craft-ai --tool claude-code --tool cursor

🔒 Security

PaperCraft is a prompt engineering project — it installs Markdown skill files and rules into your project. It does not execute code, access APIs, or send data anywhere. All skills run locally within your AI coding tool.

🇨🇳 中文说明

PaperCraft 是一个全流程 CS 科研 AI 伙伴，包含 12 个技能 + 17 条研究规则 + 16 个 CCF-A 会议画像，帮助研究者从想法到录用论文的每一步。

它解决什么问题？

| 没有 PaperCraft | 有了 PaperCraft | |:-:|:-:| | AI 搜论文返回 100 条无关结果 | /paper-search — 四层渐进式搜索 + 质量过滤 + 引用链追踪 | | 写完才发现想法不新 | /paper-idea — 动手前先做新颖性检查 | | 实验设计要 8 卡 A100，你只有 1 卡 3090 | /paper-plan — 先问你的 GPU，再在约束内设计方案 | | 论文被审稿人 2 拒了 | /paper-review — 提交前模拟 2-3 个审稿人，含会议评分标准 | | 不知道从哪开始 | /paper-launch — 一条命令从想法到完整项目方案 |

快速安装

npx paper-craft-ai           # 推荐，跨平台（Windows/macOS/Linux）

核心特色

• 会议适配 — 同一个想法，投 NeurIPS 和投 AAAI 写法完全不同，PaperCraft 自动适配
• 资源感知 — 先问 GPU 再设计实验，不会设计出跑不了的方案
• 解释先行 — 每个输出都附带"为什么"和"下一步做什么"，不再无脑输出
• 中国特供 — A800/昇腾910B 训练配方、CANN 兼容矩阵、镜像源配置
• 六大工具支持 — Claude Code / Cursor / Codex / Gemini / Copilot / Windsurf

12 个技能一览

| 技能 | 功能 | |------|------| | /paper-launch | 一键启动：想法 → 文献搜索 → 新颖性检查 → 实验方案 → 项目包 | | /paper-search | 四层渐进式搜索，质量+相关性双过滤，雪球引用追踪 | | /paper-idea | 创新点挖掘，新颖性决策树，风险评估 | | /paper-plan | GPU 优先的实验设计，开源代码/数据集搜索，消融实验规划 | | /paper-data | 数据集选择，污染检查，一键处理流水线 | | /paper-env | 一次性环境搭建，CUDA/PyTorch 兼容矩阵 | | /paper-train | 自动推导训练参数，NaN/OOM 调试决策树，结果分析 | | /paper-story | 6 种 CS 论文叙事模式，逻辑链验证 | | /paper-write | 8 个章节写作指南，会议特定风格，主张-证据对齐 | | /paper-review | 模拟 2-3 个审稿人，含"杠精"审稿人，会议评分标准 | | /paper-rebuttal | 审稿意见解析，证据驱动的回复模板 | | /paper-audit | 50+ 项检查清单，6 个维度，提交前全链路审计 |

⭐ Star History

More Tools for AI Developers

• 🔄 ContextSync — 8 rules + 3 skills that make AI respect your project conventions
• 🤖 ML Playbook — 12 rules + 4 skills that make Claude your senior ML engineer

License

BSD-3-Clause

如果 PaperCraft 帮助了你的研究，请给个 Star ⭐

Made with ❤️ by charlotte-12s